[WIP] neural biasing with early context injection and text perturbation

[huangruizhe]

This PR implements the paper: Improving Neural Biasing for Contextual Speech Recognition by Early Context Injection and Text Perturbation . This also addresses issue #1709 .

This codes can replicate the paper, but it is still a WIP -- needing codes refactoring and cleaning up.

TODO:

• Make soft links to the original standard ASR recipe -- avoid copying scripts
• Expose neural biasing and so on as an external function call to contextualize the models
• Simplify the scripts for loading contexts -- only need an example for librispeech
• Get evaluation scripts for common/rare words as in the contextual librispeech setup